Tiering, not gatekeeping

STRONG. MODERATE. SPECULATIVE. A debunk is worth exactly as much as a confirmation.

Every factual, scientific, or physiological claim in a manuscript gets tiered by evidence strength — claim-by-claim, not file-by-file — before print. A popular claim that doesn't hold up is reported as a finding, never quietly dropped because it's inconvenient.

See the tier taxonomy View on GitHub

Three tiers, applied claim-by-claim

The same source can support one tier for what it shows in a study, and a different tier for the human claim built on top of it.

STRONG

Well-replicated human evidence, or genuine consensus.

"Autophagy is a conserved cellular degradation pathway — one of the most solidly established facts in cell biology."

MODERATE

Small trials, mostly animal/mechanistic, mixed human results.

"Nutrient restriction reliably activates autophagy in animal models — strong there, moderate as a basis for human claims."

SPECULATIVE

Hypothesis, or popular claim, weak support. Debunks land here too.

"Autophagy kicks in at 16 hours of fasting" — not supported in humans even at 24 hours.

The refusal gate this skill runs on

"I won't drop a claim just because the evidence debunks it — a debunked popular claim is tagged [SPECULATIVE / debunk] and reported as a finding, exactly like a confirmed one, never quietly omitted." The mechanism checking out doesn't rescue the specific number built on top of it.

The safety file — over-cautious by design

Where a topic carries real physical risk, contraindications get their own dedicated file, separate from general research, so they can't be diluted.

TierWhat it meansExample
ABSOLUTEConsensus says "do not," not "be careful"Pregnancy, type 1 diabetes, active eating-disorder history
RELATIVEPossible, but only with clinical screening + monitoringDaily prescription medication, cardiac risk factors
The danger isn't always where you'd expect

For extended fasting, the most dangerous moment is often not the fast itself but the refeeding — reintroducing food too fast can trigger a potentially fatal electrolyte collapse. Naming that explicitly is what a safety file is for.

When a source can't be read — say so first

A blocked tool or an unreadable primary source is normal. Hiding it isn't an option.

Manuscripts verified on this discipline

Proof of shipped verification passes goes here — added by the operator, not invented.

Setup

  1. Load this folder into a Claude Project, or point a Claude Code session at it — SKILL.md lets it auto-discover and trigger from a natural request.
  2. Bring the manuscript's claims and any claim-sensitivity flags from extraction — this skill grades what's already been claimed, it doesn't research from scratch.
  3. Let it tier every claim, claim-by-claim, and produce a dedicated safety file if the topic warrants one.
  4. Any source that couldn't be read directly gets a sourcing-confidence banner before any finding, not after.
  5. Hand the tiered results to revision and to downstream content — never a looser tier than what was actually earned.

FAQ

Does it find the sources itself?

No — that's the Research worker's job. This skill tiers and flags claims and sources it's handed or pointed at.

What if a claim debunks something the book wants to say?

It gets reported anyway, tagged [SPECULATIVE / debunk] — see the refusal gate above. A debunk is a finding, not a loss.

Can it soften a safety warning if it's inconvenient for a chapter?

No. ABSOLUTE contraindications never get downgraded to RELATIVE to make a chapter read more permissive.

What happens if a fetch tool is blocked mid-research?

The file states it plainly at the top, before any finding, and tags every downstream figure as unverified until it's actually checked.